SEMINAR 2026

Abstract

Over the past century, advances from the discovery of quantum mechanics to predictive materials computation have led to the development of first-principles methods, most notably density functional theory (DFT), enabling the design and prediction of materials at the atomic and electronic levels. However, the high computational cost of these methods limits their application to large-scale, high-throughput calculations and thus constrains data-driven materials discovery. In this talk, I will introduce an emerging research direction in which deep learning is integrated with first-principles electronic structure calculations to address the computational bottlenecks of conventional methods. In particular, I will highlight DeepH, a deep-learning Hamiltonian approach developed by our group that provides a general and transferable neural-network framework for DFT. By learning DFT Hamiltonians from calculations on small-scale structures, DeepH can generalize to larger materials systems and achieve orders-of-magnitude gains in computational efficiency while retaining first-principles accuracy. I will discuss its performance across representative applications and its implications for scalable electronic-structure prediction. Looking ahead, increasingly large training datasets may enable neural-network-based approaches to evolve toward more efficient generative models for materials, opening new opportunities for AI-driven discovery in physics and materials science.

References:

• Nature Computational Science 2, 367 (2022); 3, 321 (2023); 4, 752 (2024); 5, 1133 (2025)
• Nature Communications 14, 2848 (2023); 15, 8815 (2024)
• Physical Review Letters 132, 096401 (2024); 133, 076401 (2024)
• Quantum Frontiers 3, 8 (2024); Science Bulletin 69, 2514 (2024)
• Physics, 53, 442 (2024); 54, 1 (2025) DeepH code: arXiv:2601.02938 https://github.com/mzjb/DeepH-pack